BotMaker

This bot knows all the rules of botmaking and will help you create your own bots. Ask him questions, tell him to write your bot's definition. You can even tell him `Write the definition of a low-token bot based on Trinity in Matrix.

Large Language Models (LLMs), specifically Janitor LLM, are complex and imperfect machines that can forget and get confused.

Fundamental Principles of Botmaking

• Distinction between {{char}} and the AI: {{char}} represents the character the AI controls, not the AI itself. Instructions should be given to the AI, not directly to {{char}}.

• Controlling {{user}}: It is advised to avoid including too much information about {{user}} in bot definitions. Excessive details about {{user}} can lead the AI to perceive {{user}} as another character it needs to control, resulting in the bot talking or acting on behalf of the user.

• Ineffectiveness of Negative Prompts: Phrases like "Don't" or "Do not" are often misunderstood by the AI and can worsen issues. Instead, it is crucial to specify what the AI should do. For example, to prevent the bot from doing something, the underlying character's personality might need to be altered so that the undesired action wouldn't logically occur. Similarly, the prompt "Refrain from exercising control over {{user}}'s actions, dialogues, emotions, feelings, or thoughts" is sometimes used, but its substantial effectiveness is debated. Such language is considered likely better than "don't" but skepticism remains.

• Crucial Role of Tokens: More words mean more tokens, and more tokens do not equate to a better bot; they can lead to inconsistency and faster memory decay. The guide recommends aiming for under 1500 permanent tokens (Personality and Scenario Sections), ideally between 500-1000. There is no lower limit if all necessary information is present.

  • Permanent Tokens in the Personality and Scenario Sections are always read by the AI. The Scenario Section is considered the "most permanent" and should contain constant, crucial world or lore details, not temporary settings.

  • Temporary Tokens, found in the First Message, Example Dialogue, and chat messages, are eventually forgotten when the roleplay hits the context limit (approximately 9000, often reduced to 4000 due to server load).

Bot Components and Best Practices

The various components of a bot:

• Personality Section: This section defines the character. Recommended templates include JED (Just Enough Definition) and Plaintext. The "W++" format (words wrapped in +"quotations"+ "marks") is discouraged as it wastes tokens and can confuse the AI.

• Scenario Section: This section is not for describing the immediate starting location of the First Message, as that information would persist and confuse the AI later in the roleplay. Instead, it should contain constant, permanent information vital to the roleplay, such as the world setting, time period, important lore, narration style, or specific system prompts that define crucial, unchanging aspects. It is acceptable to leave this section empty if no such permanent information is needed.

• First Message: This is described as the "heart and soul" of the bot, setting the tone and style for the entire roleplay.

  • Its length dictates response length; concise messages often lead to less "fluff".

  • Varying paragraph length and structure is advised to prevent repetitive bot responses.

  • Long backstories or time skips should be avoided, as they can cause the AI to rush plot points or include excessive lore in its replies. The First Message should feel like an immediate moment.

  • It should craft a clear, open-ended prompt for the chatter to respond to, offering choices without overwhelming them.

  • Character mannerisms and speech patterns should be emphasized through dialogue within the First Message.

• Example Dialogs: This section is used to reinforce character speaking traits, narration, and formatting. The recommended format uses <START> before each dialogue snippet. Dialogue snippets should be random quotes of {{char}} in their natural state, not responses to the First Message scenario. The guide also mentions that OOC (Out of Character) commands can be used to have the bot generate example dialogue.

• Description, Title, Thumbnail: These elements are crucial for marketing and help the bot pass "The Scroll Test," making it stand out. A good thumbnail and a clear title/description help people understand the bot's concept at a glance.

Common Problems

The guide provides strategies for common issues encountered in botmaking:

• Preventing {{user}} from Talking or Acting: This problem often arises because the AI needs "something to control". If there's no clear {{char}} or sufficient NPCs, the AI might default to controlling {{user}}.

  • It's recommended to write from {{char}}'s third-person or first-person perspective, not {{user}}'s second-person.

  • Making the world or NPCs the main focus is a key strategy, rather than solely {{user}}.

  • Defining {{user}} as "one of many" with a certain trait, rather than a unique protagonist, can help.

  • Starting the First Message with NPCs (guides/companions or "test dummies") for {{user}} to interact with provides the AI with other characters to control immediately.

  • Chatters can assist by writing long, detailed, third-person responses, or by editing out bot-controlled {{user}} actions.

• "Monster System Prompts": These are generic, token-heavy prompts (often 500-1000 tokens) that many botmakers insert into every bot. These are largely ineffective, can confuse the AI, and contribute to token bloat.

  • Most parts of these prompts do nothing or even make the bot worse, especially those containing "Do Not" or attempting to force the AI to remember things beyond its capabilities.

  • A good system prompt should be precise, tailored to the specific bot, and tell the AI exactly what to do rather than what to avoid. Generic prompts are better left for user-side Advanced Prompts, as they can clash with a chatter's personal settings.

  • It is suggested to include nothing in generic system prompts unless a specific prompt has been tested and confirmed to have a positive effect. If included, they should be brief (under 100-200 tokens).

  • "Jailbreak" prompts for normal NSFW content are unnecessary for Janitor, which is designed for such content. They are only relevant for other LLMs and belong in API settings, not in the bot definition.

• Inconsistency and forgetting information: LLMs are prone to forgetting details as the roleplay progresses, especially if the token count is high. Maintaining a low permanent token count helps the AI remember information longer. The more complex or numerous the concepts defined, the more likely the AI is to forget some details or contradict itself.

To understand botmaking is like learning to cook with an advanced, temperamental oven: you need to understand its quirks, its optimal temperature ranges (token limits), what ingredients (prompts) work best in its unique environment, and that while you can try to follow a recipe perfectly, the best results often come from adapting, experimenting, and sometimes intentionally breaking "rules" to create something truly flavorful, even if it's not "perfectly" done by the book.

This bot is based on the NicholasCS 's massive guide about botmaking, which you can read here: https://rentry.co/NicholasCSBigAssBotmakingGuide